- 187 Views
- Uploaded on
- Presentation posted in: General

Linda

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.

- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

Developed at Yale University by D. Gelernter, N. Carriero et al.

Goal: simple language-and machine-independent model for parallel programming

Model can be embedded in host language

C/Linda

Modula-2/Linda

Prolog/Linda

Lisp/Linda

Fortran/Linda

Sequential constructs

Same as in base language

Parallelism

Sequential processes

Communication

Tuple Space

Tuple Space

- box with tuples (records)
- shared among all processes
- addresses associatively (by contents)

Atomic operations on Tuple Space (TS)

OUTadds a tuple to TS

READreads a tuple (blocking)

INreads and deletes a tuple

Each operation provides a mixture of

- actual parameters (values)
- formal parameters (typed variables)

Example

age: integer;

married: Boolean;

OUT ("Jones", 31, true)

READ ("Jones", ? &age, ? &married)

IN ("smith", ? &age, false)

Tuples cannot be modified while they are in TS

Modifying a tuple:

IN ("smith", ? &age, ? &married)/* delete tuple */

OUT ("smith", age+1, married)/* increment age */

The assignment is atomic

Concurrent READ or IN will block while tuple is away

- Data structure that can be accessed simultaneously by different processes
- Correct synchronization due to atomic TS operations
- Contrast with centralized manager approach, where each processor encapsulates local data

Popular programming style, also called task-farming

Collection of P identical workers, one per CPU

Worker repeatedly gets work and executes it

In Linda, work is represented by distributed data structure in TS

Scales transparently - can use any number of workers

Eliminates context switching

Automatic load balancing

- Source matrices:("A", 1, A's first row)
("A", 2, A's second row)

...

("B", 1, B's first column)

("B", 2, B's second column)

- Job distribution: index of next element of C to compute
("Next", 1)

- Result matrix:("C", 1, 1, C[1,1])
...

("C", N, N, C[N,N])

repeat

in ("Next", ? &NextElem);

if NextElem < N*N then

out ("Next", NextElem + 1)

i = (NextElem - 1)/N + 1

j = (NextElem - 1)%N + 1

read ("A", i, ? &row)

read ("B", j, ? &col)

out ("C", i, j, DotProduct (row,col))

end

Use replicated workers parallelism

A master process generates work

The workers execute the work

Work is stored in a FIFO job-queue

Also need to implement the global bound

Use a tuple representing global minimum

Initialize;

out ("min", maxint)

Atomically update minimum with new value:

in ("min", ? & oldvalue);

value = minimum (oldvalue, newvalue)

out ("min", value)

Read current minimum:

read ("min", ? &value)

Add a job:

- In ("tail", ? &tail)
- Out ("tail", tail+1)
- Out ("JQ", job, tail+1)
Get a job:

- In ("head", ? &head)
- Out ("head", head+1)
- In ("JQ", ? &job, head)

tsp (int hops, int len, int path[])

int e,me;

rd("minimum",? &min); /* update min */

if (len < min)

if (hops == (NRTOWNS-1))

in("minimum", ? &min);

min = minimum(len,min);

out("minimum",min);

else

me = path[hops];

for (e=0; e < NRTOWNS; e++)

if (!present(e,hops,path))

path[hops+1] = e;

tsp(hops+1,len+distance[me][e], path);

int min;

LINDA_BLOCK PATH;

worker()

int hops, len, head;

int path[MAXTOWNS];

PATH.data = path;

for (;;)

in("head", ? &head)

out("head", head+1)

in("job", ?&hops, ?&len, ?&PATH, head)

tsp(hops,len,path);

master(int hops,int Len, int path[])

int e,me;

if (hops == MAXHOPS)

PATH.size = hops + 1; PATH.data = path;

in("tail", ? &tail)

out("tail", tail+1)

out("job", hops, len, PATH, tail+1)

else

me = path[hops];

for (e=0; e < NRTOWNS; e++)

if (!present(e,hops,path))

path[hops+1] = e;

master(hops+1, len+distance[me][e], path);

Communication is uncoupled from processes

The model is machine-independent

The model is very simple

Possible efficiency problems

Associative addressing

Distribution of tuples

Linda has been implemented on

- Shared-memory multiprocessors (Encore, Sequent, VU Tadpole)
- Distributed-memory machines (S/Net, workstations on Ethernet)
Main problems in the implementation

- Avoid exhaustive search (optimize associative addressing)
- Potential lack of shared memory (optimize communication)

Linda preprocessor

Analyzes all operations on Tuple Space in the program

Decides how to implement each tuple

Runtime kernel

Runtime routines for implementing TS operations

Partition all TS operations in disjoint sets

Tuples produced/consumed by one set cannot be produced/consumed by operations in other sets

OUT("hello", 12);

will never match

IN("hello", 14); constants don't match

IN("hello", 12, 4); number of arguments doesn't match

IN("hello", ? &aFloat); types don't match

Based on usage patterns in entire program

Use most efficient data structure

- Queue
- Hash table
- Private hash table
- List

OUT ("foo", i);

IN ("foo", ? &j);

- First field is always constant -> can be removed by the compiler
- Second field of IN is always formal -> no runtime matching required

OUT ("vector", i, j);

IN ("vector", k, ? &l);

First and third field same as in previous example

Second field requires runtime matching

Second field always is actual -> use hash table

OUT ("element", i, j);

IN ("element", k, ? &j);

RD ("element", ? &k, ? &j);

Second field is sometimes formal, sometimes actual

If actual -> use hashing

If formal -> search (private) hash table

OUT ("element", ? &j);

Only occurs if OUT has a formal argument -> use exhaustive search

Shared-memory kernels

- Store Tuple Space data structures in shared memory
- Protect them with locks
Distributed-memory (network) kernels

- How to represent Tuple Space?
- Several alternatives:
Hash-based schemes

Uniform distribution schemes

Each tuple is stored on a single machine, as determined by a hash function on:

1. search key field (if it exists), or

2. class number

Most interactions involve 3 machines

P1: OUT (t); -> send t to P3

P2: IN (t); -> get t from P3

Network with reliable broadcasting (S/Net)

broadcast all OUT’s ->) replicate entire Tuple Space

RD done locally

IN: find tuple locally, then broadcast

Network without reliable broadcasting (Ethernet)

OUT is done locally

To find tuple (IN/RD), repeatedly broadcast

Performance is hard to predict, depends on

Implementation strategy

Application

Example: global bound in TSP

Value is read (say) 1,000,000 times and changed 10 times

Replicated Tuple Space -> 10 broadcasts

Other strategies -> 1,000,000 messages

Multiple TS operations needed for 1 logical operation

Enqueue and dequeue on shared queue each take 3 operations

+ Very simple model

+ Distributed data structures

+ Can be used with any existing base language

- Tuple space operations are low-level

- Implementation is complicated

- Performance is hard to predict